1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

use crate::infer::*;
use crate::internal::*;

// TODO tricky to re-express in "core" because of the multiple hot point... do
// we need one more reduce ?
#[derive(Debug, Clone, new, Default, Hash)]
pub struct LayerHardmax {
    axis: isize,
    coerce_to_2d: bool,
}



impl Expansion for LayerHardmax {
    fn name(&self) -> Cow<str> {
        "LayerHardmax".into()
    }

    fn rules<'r, 'p: 'r, 's: 'r>(
        &'s self,
        solver: &mut Solver<'r>,
        inputs: &'p [TensorProxy],
        outputs: &'p [TensorProxy],
    ) -> InferenceResult {
        rules(solver, inputs, outputs)
    }

    fn wire(
        &self,
        name: &str,
        target: &mut TypedModel,
        inputs: &[OutletId],
    ) -> TractResult<TVec<OutletId>> {
        use tract_core::ops::{array, change_axes, nn};
        let input = inputs[0];
        let input_fact = target.outlet_fact(input)?.clone();
        let input_dt = input_fact.datum_type;
        let rank = input_fact.rank();
        let axis = if self.axis < 0 { rank as isize + self.axis } else { self.axis } as usize;
        let suffix_dim: TDim = input_fact.shape[axis..].iter().product();
        let dim = if self.coerce_to_2d {
            suffix_dim.to_usize()
        } else {
            input_fact.shape[axis].to_usize()
        }
        .context("Assumes known dimension on working axes suffix.")?;
        let off = tensor0(0f32).cast_to_dt(input_dt)?.into_owned().into_arc_tensor();
        let on = tensor0(1f32).cast_to_dt(input_dt)?.into_owned().into_arc_tensor();
        let mut wires = inputs.into();
        if self.coerce_to_2d {
            wires = target.wire_node(
                format!("{name}.reshaped"),
                AxisOp::Reshape(axis, input_fact.shape[axis..].into(), tvec!(suffix_dim.clone())),
                &[input],
            )?;
        }
        wires = target.wire_node(
            format!("{name}.argmax"),
            nn::Reduce::new(tvec!(axis), nn::Reducer::ArgMax(false)),
            &wires,
        )?;
        wires =
            target.wire_node(format!("{name}.rm_axis"), change_axes::AxisOp::Rm(axis), &wires)?;
        wires = target.wire_node(
            format!("{name}.hardmax"),
            array::OneHot { axis, dim, off, on },
            &wires,
        )?;
        if self.coerce_to_2d {
            wires = target.wire_node(
                format!("{name}.hardmax_reshaped"),
                AxisOp::Reshape(axis, tvec!(suffix_dim), input_fact.shape[axis..].into()),
                &wires,
            )?;
        }
        Ok(wires)
    }
}

#[derive(Debug, Clone, new, Default, Hash)]
pub struct LayerLogSoftmax {
    axis: isize,
    coerce_to_2d: bool,
}



impl Expansion for LayerLogSoftmax {
    fn name(&self) -> Cow<str> {
        "LayerLogSoftmax".into()
    }


    fn rules<'r, 'p: 'r, 's: 'r>(
        &'s self,
        solver: &mut Solver<'r>,
        inputs: &'p [TensorProxy],
        outputs: &'p [TensorProxy],
    ) -> InferenceResult {
        rules(solver, inputs, outputs)
    }

    fn wire(
        &self,
        name: &str,
        target: &mut TypedModel,
        inputs: &[OutletId],
    ) -> TractResult<TVec<OutletId>> {
        let softmax = LayerSoftmax { axis: self.axis, coerce_to_2d: self.coerce_to_2d }
            .wire(name, target, inputs)?;
        target.wire_node(format!("{name}.logsoftmax"), tract_core::ops::math::ln(), &softmax)
    }
}

#[derive(Debug, Clone, new, Default, Hash)]
pub struct LayerSoftmax {
    axis: isize,
    coerce_to_2d: bool,
}



impl Expansion for LayerSoftmax {
    fn name(&self) -> Cow<str> {
        "LayerSoftmax".into()
    }

    fn rules<'r, 'p: 'r, 's: 'r>(
        &'s self,
        solver: &mut Solver<'r>,
        inputs: &'p [TensorProxy],
        outputs: &'p [TensorProxy],
    ) -> InferenceResult {
        rules(solver, inputs, outputs)
    }

    fn wire(
        &self,
        name: &str,
        target: &mut TypedModel,
        inputs: &[OutletId],
    ) -> TractResult<TVec<OutletId>> {
        let input = inputs[0];
        let rank = target.outlet_fact(input)?.rank();
        let dt = target.outlet_fact(input)?.datum_type;
        let axis = if self.axis < 0 { rank as isize + self.axis } else { self.axis } as usize;
        let reducing_axes =
            if self.coerce_to_2d { (axis..rank).collect::<TVec<usize>>() } else { tvec!(axis) };
        target.wire_node(name, tract_core::ops::nn::Softmax::new(reducing_axes, dt), inputs)
    }
}

fn rules<'r, 'p: 'r, 's: 'r>(
    s: &mut Solver<'r>,
    inputs: &'p [TensorProxy],
    outputs: &'p [TensorProxy],
) -> InferenceResult {
    check_output_arity(outputs, 1)?;
    s.equals(&outputs[0].datum_type, &inputs[0].datum_type)?;
    s.equals(&outputs[0].rank, &inputs[0].rank)?;
    s.equals(&outputs[0].shape, &inputs[0].shape)?;
    Ok(())
}